GraphRAG
Properties
created
Empty
modified
26.07.2026, 13:49
published
Empty
sources
Empty
topics
GraphRAG
authors
Jakub
ai-assisted
Yes
- GraphRAG represents new approach for constructing the knowledge base for the RAG application. It can improve the quality of the output of the RAG application. But this come at a significant cost. The goal of this task is to explore whether the improvement in quality justifies the cost and whether we can use this approach in our projects.
# Approaches
- NaviRAG: Reframes RAG as active knowledge navigation; builds offline Knowledge Trees and navigates them top-down at query time, outperforming GraphRAG and HippoRAG2 on cross-region evidence tasks
- HippoRAG 2: Builds a flat OpenIE-based KG over passages and applies Personalized PageRank at query time for associative, multi-hop retrieval
- G-Retriever: Retrieves a connected subgraph via Prize-Collecting Steiner Tree, then feeds it to the LLM as both text and a GNN-encoded soft prompt
- LightRAG: Flat entity-relation KG with dual-level (low/high) index keys; dual vector lookup + 1-hop graph expansion at query time; supports incremental updates by graph union
- GraphRAG: initial GraphRAG paper
- An Ontology-Driven Graph RAG for Legal Norms: A Structural, Temporal, and Deterministic Approach
- Building ontologies for the Legal domain
- Law GraphRAG: An Advanced Legal QuestionAnswering System
# Resources
- Papers:
- Graph Retrieval-Augmented Generation: A Survey
- https://dl.acm.org/doi/10.1145/3777378
- https://github.com/pengboci/GraphRAG-Survey
- The authors formalize the GraphRAG workflow into three core stages: Graph-Based Indexing (constructing and organizing graph databases from open knowledge graphs or self-constructed data), Graph-Guided Retrieval (extracting relevant graph elements like nodes, triplets, paths, or subgraphs using non-parametric, LM-based, or GNN-based retrievers), and Graph-Enhanced Generation (converting retrieved graph structures into formats processable by language models). The survey systematically categorizes existing methodologies across these stages, examines retrieval paradigms (once, iterative, and multi-stage), discusses training strategies, and analyzes downstream applications spanning question answering, recommendation systems, and domain-specific tasks in healthcare, finance, and e-commerce.
- HybridRAG
- GraphRAG
- Graph Retrieval-Augmented Generation: A Survey
- Existing tools and frameworks for leveraging knowledge graphs in RAG:
- GraphRAG Website
- Good overview video of GraphRAG
- GraphRAG course
# Domain-Specific GraphRAGs
- Legal
- An Ontology-Driven Graph RAG for Legal Norms
- JURIX 2025 — ontology-grounded, structure-aware temporal graph for versioned legal text; deterministic point-in-time retrieval and provenance reconstruction
- Github:
hmartim/sat-graph-api
- 10 stars
- LegalGraphRAG
- ACL 2026 — three-layer Hierarchical Legal Graph (Fact / Ontology / Rule) with a Researcher → Auditor → Adjudicator multi-agent pipeline for legal judgment prediction
- Github: Zovi343/LegalGraphRAG
- Agentic Knowledge Graph Construction.
- An Ontology-Driven Graph RAG for Legal Norms